TileLang
MIT Pythonic DSL for GPU/CPU/NPU kernels on TVM
Python / Python · OSS/MIT · 8.6k
Reviewed 2026-10-03 · last push 2026-10-19 · repo checked 2026-10-19
What's Good
MIT (tile-ai/tilelang). PyPI `tilelang`; docs and install at tilelang.com. Multi-backend dialect (CUDA/ROCm/Metal/LLVM/Ascend 950), LSP, puzzles, and active kernel examples (FlashAttention, block-scaled MMA). Free path is pip + your own accelerators — no SaaS seat. ★~8,258 / F831.
The Catch
Not an agent product. Serious kernels need real GPU/NPU hardware and TVM-shaped mental models; CPU paths will not match CUDA blog numbers. Vs NVIDIA Model Optimizer: ModelOpt quantizes/prunes/distills HF models for TensorRT-LLM/vLLM; TileLang authors the kernels themselves. v0.1.x still drops legacy APIs — pin a release before production kernels.
Verdict
MIT Pythonic kernel DSL when you write GPU/NPU tiles on TVM — not ModelOpt quantization; bring the hardware.
Embed
Maintainer? Add the badge to your README. It is free, links to this review, and there is no paid version.
Markdown
[](https://stackgems.com/gems/tilelang)
HTML
<a href="https://stackgems.com/gems/tilelang"><img src="https://stackgems.com/badge/tilelang.svg" alt="TileLang reviewed on Stack Gems" height="20"></a>
Related
Python →- marimoApache-2.0 reactive Python notebooks as git-friendly scripts
- SupervisionModel-agnostic computer-vision toolkit for detections, zones, and video
- HishelBSD-3 elegant HTTP caching for Python
- NVIDIA Model OptimizerApache-2.0 quantization/pruning/NAS/distill for HF → TensorRT-LLM/vLLM
- Data FormulatorMIT AI-powered data exploration and visualization (Microsoft)